Direct-LLM: A Direct Interface Paradigm

• Direct-LLM is a paradigm that bypasses traditional intermediary layers in LLM pipelines to enhance speed, accuracy, and control.
• It facilitates direct decision generation, multilingual inference, and preference-based alignment by reducing external orchestration.
• Direct-LLM methods offer practical performance gains but may shift complexity to internal reasoning and runtime optimization.

Direct-LLM is a recurring but non-equivalent label in recent research for architectures in which a LLM is placed in a more immediate relation to the target task, representation, or execution substrate than in mediated pipelines. Across the cited works, the omitted intermediary varies: a solver, reward model, translation layer, text serialization stage, code-summary scaffold, retrieval grounding constraint, or filesystem/page-cache path. The term therefore denotes a family of “direct” interfaces rather than a single standardized method, spanning direct constrained optimization, direct preference-based alignment, direct multilingual inference, direct latent-state synthesis, direct speech-to-speech translation, direct prompting for code modification, direct manipulation of the LLM core, and direct hardware access for inference (Tso et al., 25 Feb 2026, Liu et al., 2024, Intrator et al., 2024, Liu et al., 12 Jun 2026, Arya et al., 22 Jan 2026, Lin et al., 28 Apr 2026).

1. Terminological scope and recurrent design pattern

Across these works, “direct” consistently marks the removal of an intermediate layer that had previously structured, constrained, or translated the model’s interaction with the task. In operations research, it means outputting decision variables instead of solver code; in multilingual inference, answering in the source language instead of routing through English; in agentic synthesis, consuming KV caches instead of concatenated branch text; in alignment, optimizing from preference data without reward-model training and PPO; in systems work, reading remote memory or NVMe-backed KV cache without HBM staging or filesystem/page-cache mediation (Tso et al., 25 Feb 2026, Liu et al., 2024, Intrator et al., 2024, Liu et al., 12 Jun 2026, Lin et al., 28 Apr 2026, Jeong et al., 29 Apr 2026).

A plausible unifying view is that Direct-LLM methods move the LLM closer to the “native” object of interest—decisions, preferences, latent states, speech, design intent, or storage-resident tensors—while correspondingly reducing external orchestration. This suggests a common trade-off: directness can remove latency, error propagation, and interface loss, but it can also shift burden onto the model’s internal reasoning or onto carefully engineered runtime support.

Usage of “Direct-LLM”	Removed intermediary	Representative paper
Direct decision output	Solver formulation and external optimizer	(Tso et al., 25 Feb 2026)
Direct policy alignment	Reward/cost models and PPO	(Liu et al., 2024)
Direct multilingual inference	Pre-translation through English	(Intrator et al., 2024)
Direct latent synthesis	Text serialization of branch outputs	(Liu et al., 12 Jun 2026)
Direct understanding manipulation	Prompt-only steering of hidden task structure	(Zhang et al., 5 Aug 2025)
Direct hardware access	HBM staging or filesystem/page-cache path	(Lin et al., 28 Apr 2026)

2. Direct solution generation and decision making

In constrained optimization, “Direct-LLM” is used in the most literal solver-like sense. “ConstraintBench” defines it as using an LLM as a solver that directly outputs decision variables for a fully specified constrained optimization problem, without generating solver code or calling an external optimizer (Tso et al., 25 Feb 2026). The benchmark covers 10 operations research domains, 20 tasks per domain, hence 200 tasks total, with natural-language scenarios, domain-specific JSON schemas, deterministic verification, and Gurobi-proven optima. The underlying tasks are formulated as MIPs, and evaluation separates feasibility from optimality via a solver-referenced gap criterion with a 0.1% threshold (Tso et al., 25 Feb 2026).

The empirical picture is sharply asymmetric. Across 200 tasks and six frontier models, the best feasibility rate is 65.0%, yet feasible solutions average 89 to 96% of the Gurobi-optimal objective; no model exceeds 30.5% on joint feasibility and optimality within 0.1% of the solver reference (Tso et al., 25 Feb 2026). Domain difficulty varies widely: average feasibility ranges from 83.3% in production mix to 0.8% in crew assignment, and the reported failure modes include duration constraint misunderstanding, entity hallucination, and a feasibility–optimality decoupling in facility location and vehicle routing where models achieve high feasibility but 0% optimality (Tso et al., 25 Feb 2026). The result is a precise characterization of direct decision generation as a strong heuristic capability constrained primarily by feasibility, not by objective improvement once inside the feasible region.

A narrower but related usage appears in multi-phase budget allocation. “ZEBRA” defines “LLM-direct” as asking the controller, given the same pipeline description and total budget, to output a per-phase split directly, without utility-curve estimation or an explicit solver (Hamri et al., 19 May 2026). On a 150-task APPS coding benchmark, both ZEBRA variants outperform LLM-direct on every aggregate metric; at a budget of $\alpha = 0.5$ of the unconstrained spend, ZEBRA recovers 94.4% of unconstrained quality versus 88.1% for LLM-direct, and on a 3-phase HotpotQA pipeline ZEBRA beats LLM-direct by 14.3 percentage points (Hamri et al., 19 May 2026). In both papers, directness means bypassing formal optimization machinery; the difference is that ConstraintBench measures the intrinsic solver-like competence of the model, whereas ZEBRA shows that directly asking an LLM to solve a continuous allocation problem is inferior to using the LLM only to estimate curve parameters and then solving the resulting knapsack instance algorithmically.

3. Direct preference-based alignment

In alignment research, “direct” refers to optimizing the policy from preferences without explicit reward-model fitting and RL. The canonical formulation is Direct Preference Optimization, in which the model is trained from pairwise preferences using the reference-policy-relative log-probability margin

$$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$

rather than through a separately learned reward and PPO loop (Liu et al., 2024). “Constrained DPO” extends this to safe alignment by introducing a safety cost constraint, dual variable updates, and a combined preference function $r_\lambda = r - \lambda c$ (Liu et al., 2024). On BEAVERTAILS with threshold $C_{limit}=0$, vanilla DPO achieves reward $6.16 \pm 6.15$ and cost $12.45 \pm 18.63$, whereas C-DPO with $\lambda=0.4$ attains reward $4.26 \pm 5.59$ and cost $-0.58 \pm 16.70$, and Safe-RLHF’s Beaver-v1 attains reward $0.39 \pm 6.01$ and cost $$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$0 (Liu et al., 2024). The paper’s central claim is therefore not merely that direct optimization is cheaper, but that it can realize a nearly optimal helpfulness–harmlessness trade-off under an explicit constraint.

“Primal-Dual Direct Preference Optimization for Constrained LLM Alignment” generalizes this line by first training a standard DPO policy on reward preference data and then using a rearranged Lagrangian DPO objective on cost preference data, together with projected dual updates, to handle constrained alignment without training reward and cost models or requiring prior knowledge about the optimal solution (Du et al., 7 Oct 2025). The paper establishes explicit bounds on suboptimality and constraint violation, and extends the method to an online setting with exploration bonuses to remove dependence on fixed preference-data coverage (Du et al., 7 Oct 2025). This is one of the clearest formalizations of Direct-LLM as a replacement for model-based RLHF machinery.

A different direct-alignment variant is “Direct LLM Alignment Through Self-Rewarding Contrastive Prompt Distillation,” which generates its own preference data by sampling under contrastive prompt pairs, scores those pairs by a self-rewarding probability ratio, and then applies DPO with those scores (Liu et al., 2024). Its central claim is that DLMA can surpass RLHF without relying on human-annotated preference data (Liu et al., 2024). In a task-specific application, multilingual counterspeech generation uses SFT plus DPO to align Llama-3 on hate-speech/counter-narrative pairs with GPT-4o-generated rejected outputs; on English, the DPO-aligned run reaches JudgeLM 1191.0 versus 990.0 for Llama-3 instruct SFT and 965.5 for Llama-3 base SFT, while also improving Basque and Italian metrics such as ROUGE-L and BERTScore (Wadhwa et al., 2024). Here, directness refers not to the task domain but to the alignment mechanism: preference supervision acts on the policy itself rather than through an external evaluator.

4. Direct interfaces to latent, speech, and multilingual representations

A second major sense of the term concerns direct access to representations that are normally serialized or translated away. “Parallel-Synthesis” addresses agent DAG workflows in which multiple worker branches produce outputs that are normally flattened into a single text prefix. Its proposal is to let a synthesizer consume workers’ KV caches directly, using RoPE re-encoding, a cache mapper, and synthesizer-side LoRA adapters (Liu et al., 12 Jun 2026). Across nine downstream datasets spanning math, code, science QA, GAIA, and multi-agent database diagnosis, Parallel-Synthesis matches or outperforms text-based synthesis on seven datasets, remains close on the other two, and reduces time-to-first-token by 2.5x–11x (Liu et al., 12 Jun 2026). The system therefore treats cache states as a first-class inter-agent communication medium rather than a by-product of token generation.

In multilingual language use, “direct inference” refers to bypassing English as an internal pivot. The PaLM2 study compares direct inference in the source language against pre-translation into English across 108 languages and six benchmarks, including both closed-ended and open-ended tasks (Intrator et al., 2024). The strongest reported result is that PaLM2-L consistently outperforms pre-translation in 94 out of 108 languages, with direct inference favored in 90.9% of languages on XCOPA, 100.0% on XStoryCloze, and 95.2% on BeleBele; the advantage persists even when open-ended predictions and references are translated into English for evaluation (Intrator et al., 2024). The directness here is semantic and linguistic: the model reasons in the user’s language rather than through an auxiliary translation layer that can introduce noise and break extractive alignment.

An even stronger modality-level version appears in direct speech-to-speech translation. DS2ST-LM integrates a Whisper speech encoder, a learnable projection module, a Qwen2-0.5B LLM, and a timbre-controlled vocoder in a single-stage framework, constructs the 1000-hour bilingual corpus GigaS2S-1000, and studies both speech-derived and text-derived semantic tokens (Arya et al., 22 Jan 2026). Speech-derived S3 tokens outperform text-derived tokens; the simple Linear projector outperforms Conv1D-Linear and Q-Former despite slower convergence; and the system improves over cascaded and ST+TTS baselines across lexical and semantic metrics while also achieving speaker similarity 0.83 and DNSMOS 3.54 on fr–en (Arya et al., 22 Jan 2026). In this setting, the LLM is directly conditioned on projected speech features and directly emits semantic tokens for speech synthesis rather than serving as a text MT component inside a cascade.

5. Direct manipulation of the LLM core and of externalized understanding

In security work, “Direct-LLM” can mean attacking the LLM core itself rather than the retriever or tools around it. “Targeting the Core” shows that a short adversarial prefix such as “Ignore the document” can cause RAG-based agents to disregard retrieved context and produce unsafe outputs, with high attack success rates under adaptive and ArtPrompt conditions (Li et al., 2024). The quantitative effect is stark: Mistral-7B without defenses reaches baseline ASR 0.661, adaptive ASR 0.925, and ArtPrompt ASR 0.705, while Llama3.1 without defenses reaches 0.054, 0.706, and 0.696 respectively (Li et al., 2024). The directness here is not a capability but a vulnerability: the attacker writes directly into the instruction stream seen by the model.

In software engineering, “Direct Instruction Prompting” denotes free-form natural-language requests for code modification, contrasted with summary-mediated prompting (Tang et al., 2 Aug 2025). In an exploratory study with 15 developers, direct-only prompting accounts for 42.2% of prompt usage and yields 84.2% success, whereas summary-only prompting yields 93.0% success; direct prompting is rated easier to specify, with median 6.0 versus 5.0 on the corresponding Likert item ($$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$1), but summaries are rated more helpful for code comprehension, with median 6.0 versus 4.0 ($$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$2) (Tang et al., 2 Aug 2025). The distinction shows that directness can increase immediacy and flexibility while reducing scaffolding and control.

“NeuroSync” pushes this further by externalizing the model’s inferred coding-task structure itself. It defines LLM understanding as the tasks and their relationships implicitly encoded in the code the model is expected to generate, and represents this as a triple of intent tree, understanding graph, and mapping (Zhang et al., 5 Aug 2025). A distilled SLM approximates a two-stage teacher extractor, after which users can directly inspect and edit the graph before code generation. In the user study ($$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$3), NeuroSync reduces average task time from 23.8 to 13.9 minutes, reduces average LLM calls from 3.9 to 1.3, and lowers NASA-TLX workload from 13.26 to 8.93 (Zhang et al., 5 Aug 2025). This is a particularly strong form of Direct-LLM: the user no longer steers the model only through prompts, but directly modifies the model’s externalized intermediate understanding.

A comparable shift appears in CAD. “Semantic Direct Modeling” uses Llama 3 with CAD-specific chain-of-thought prompting to interpret natural-language design intent, map it to geometric features through a conditional feature-recognition module, and generate Siemens NX API calls for direct modeling operations (Zou et al., 5 Apr 2025). Simple prompting yields 15% overall structured-output accuracy, whereas CAD-specific CoT prompts raise this to 97.5%; the generative feature recognizer reaches IoU 99.54% and exact match 99.20% on MFCAD (Zou et al., 5 Apr 2025). Here directness means lifting interaction from vertices, edges, and faces to high-level design semantics while still executing direct geometric modifications.

6. Direct hardware access, evaluation methodology, and limits of directness

At the systems layer, the term denotes direct access to memory or storage tiers that were previously reached only through staging or OS-mediated I/O. “DAK” argues that memory offloading frameworks should not prefetch remote data into HBM before kernels can use it; instead, they should let the GPU compute directly on remote-resident weights and KV cache by repurposing the Tensor Memory Accelerator to fetch into SMEM (Lin et al., 28 Apr 2026). DAK adds an operation-wise greedy offloading algorithm, active congestion control, and TMA multicast, and reports up to 3x performance gains on NVLink-C2C and 1.8x on PCIe systems relative to prefetch-based baselines (Lin et al., 28 Apr 2026). “DUAL-BLADE” makes an analogous argument for edge KV-cache offloading: it dynamically splits tensors between a page-cache path and an NVMe-direct path, maps direct-path tensors to contiguous LBA regions, and overlaps storage I/O with GPU DMA using adaptive pipeline parallelism (Jeong et al., 29 Apr 2026). The reported gains are up to 33.1% lower prefill latency, up to 42.4% lower decode latency, and 2.2x higher SSD utilization (Jeong et al., 29 Apr 2026).

The term also appears in evaluation methodology. In syntactic assessment, “Direct Minimal Pair Analysis” argues that direct minimal-pair “wh-effect” comparisons are diagnostically clearer than Difference-in-Differences metrics for parasitic gaps (Pistotti et al., 7 Oct 2025). Using a full 8-permutation paradigm, the paper reports that GPT-2 succeeds across all four tested conditions, with mean effects of $$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$4, $$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$5, $$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$6, and $$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$7 bits and significance levels at $$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$8 or $$L_{\mathrm{DPO}}(\pi_\theta;\pi_{\mathrm{ref}}) = -\mathbb{E}\Big[\log \sigma\big(\beta \log \tfrac{\pi_\theta(y_\omega|x)}{\pi_{\mathrm{ref}}(y_\omega|x)} - \beta \log \tfrac{\pi_\theta(y_\ell|x)}{\pi_{\mathrm{ref}}(y_\ell|x)}\big)\Big],$$9, thereby attributing prior ambiguity partly to metric choice rather than to lack of syntactic generalization (Pistotti et al., 7 Oct 2025). Directness here concerns how the model is interrogated: the evaluation manipulates one syntactic variable while holding the critical region fixed.

The limits of directness are equally explicit in software-model querying. In the automotive setting, direct full-context prompting supplies the entire Ecore model to the LLM in one prompt, whereas the agentic alternative uses file tools and incremental access; the paper finds accuracy to be comparable, but the agentic approach is significantly more efficient in token usage and, for large software models in automotive practice, “the only viable solution” (Mazur et al., 16 Jun 2025). This establishes a general boundary condition. Direct-LLM methods are often attractive when mediation introduces unnecessary translation, latency, or error propagation, but they are not uniformly dominant. ConstraintBench shows that direct solution generation can fail on feasibility (Tso et al., 25 Feb 2026); direct core access can be a security liability in RAG (Li et al., 2024); and direct full-context prompting can become impractical at scale (Mazur et al., 16 Jun 2025). Conversely, direct multilingual inference, direct latent synthesis, direct speech-to-speech translation, and direct hardware access can outperform mediated baselines when the removed intermediary is itself the main source of loss or overhead (Intrator et al., 2024, Liu et al., 12 Jun 2026, Arya et al., 22 Jan 2026, Lin et al., 28 Apr 2026). The literature therefore treats Direct-LLM less as a universal prescription than as a design principle whose value depends on which layer is being bypassed and which burdens that bypass transfers back onto the model or the runtime.